ElementaryCQT: A New Dataset and its Deep Learning Analysis for 2D Geometric Shape Recognition

Geometry problems at the elementary school level typically include a figure and input text, which requires interpreting both and finding a solution. Figures may include shape, side values, angles, perpendicular symbols, and other information that may or may not be mentioned in the question text. The initial step in tackling such challenges is identifying the geometric shape. Typical shapes include circles, triangles, squares, rectangles, parallelograms, trapeziums, rhombus, etc. The existing 2-dimensional (2D) shape dataset consists of plain shapes, i.e. shapes without geometrical properties. This paper presents a comprehensive dataset, ElementaryCQT (CQT stands for Circle, Quadrilateral, Triangle), of geometric 2D shapes with and without geometric properties like diagonals, perpendiculars, chords, tangents, etc. This dataset is a benchmark for machine learning and deep learning shape identification models for geometry problems. It is trained and tested with sequential CNN, VGG16, ResNet50, Inception-v3, and Inception-ResNet-v2 pre-trained deep learning models to prove their efficacy. Each network is fine-tuned on the dataset and evaluated using 10-fold cross-validation. The ResNet50 model has the best F1 score of 81.86% and accuracy of 81.97%. We performed three ensembling strategies namely, average, weighted average, and majority voting which give an F1 score of 84.07%, 84.3%, and 83.8% respectively.